Known conventionally are apparatuses for recognizing a character string printed on a surface of a medium such as a check. For example, Patent Document 1 discloses an apparatus for character recognition that includes: character segmentation means with which a character part is segmented along a circumscribing rectangular frame after detecting the character part from binarized image data; characteristic vector extraction means for extracting characteristic vectors from a character pattern obtained with the character segmentation means; characteristics storage means for storing characteristics data for each character in advance; similarity calculation means for calculating similarity between the characteristic vectors of the entered character pattern and the standard vectors of each character out of the characteristic dictionary; candidate character selection means for selecting a candidate character(s) according to the level of similarity calculated with the similarity calculation means, similar character discrimination means for further recognizing a candidate character when there exist a plurality of candidate characters selected with the candidate character selection means. Such an apparatus for character recognition including those means described above ensures appropriate recognition of a character string printed on a surface of a medium.
Meanwhile, the binarized image data described above is captured by an optical sensor for optically reading a character string printed on the surface of the medium. Specifically, a CCD line sensor (one-dimensional image pickup device) that optically reads a character string in one dimension is mostly used (Refer to Patent Document 2) from the viewpoint of reducing the manufacturing cost. In this case, main scanning for the character string on the medium is carried out by the CCD line sensor while sub-scanning is carried out through transferring the medium by manual operation or mechanical driving. Through interaction between the CCD line sensor and operation of transferring the medium, two-dimensional image data of the character string is captured, and then the two-dimensional image data is binarized to obtain the binarized image data described above.
Furthermore, Patent Document 3 for example discloses a method for character recognition in which appropriate character segmentation can be carried out through character recognition processing by selecting a most likely candidate according to a recognition data verifying score and/or linguistic likelihood even when the characters are not placed at regular intervals.
Patent Document 1
Japanese Unexamined Patent Publication (Kokai) No. JP5-120483
Patent Document 2
Japanese Unexamined Patent Publication (Kokai) No. JP5-73722
Patent Document 3
Japanese Unexamined Patent Publication (Kokai) No. JP2004-46723
However, unfortunately the apparatus for character recognition disclosed in Patent Document 2 may cause a decrease in recognition accuracy.
Specifically, when a medium transfer speed becomes so uncertain due to manual transfer operation (i.e., swiping motion in the case of a swiping-type card reader) as to cause distortion of an image in the medium transfer direction, other distortion of the image may also appear in a direction perpendicular to the medium transfer direction as the transfer operation proceed with (maybe making the image tilted and/or warped). Moreover, detachment of a lower edge part of the medium from a transfer path of the image pickup apparatus (the apparatus for character recognition) may also result in distortion of the image in the direction perpendicular to the medium transfer direction. In such a case, since the image distortion appears in the direction perpendicular to the medium transfer direction, correcting the image is difficult in general. For example, the apparatus for character recognition disclosed in Patent Document 2 enables the distortion of the image to be corrected appropriately as far as the image distortion appears in the medium transfer direction, but any image distortion in the direction perpendicular to the medium transfer direction cannot be corrected easily. As a result, character segmentation using the distorted image, as it is, leads to a decrease in recognition accuracy.
From this viewpoint, enabling appropriate character segmentation even when the characters in an array are not placed at regular intervals, the method for character recognition disclosed in Patent Document 3 may be expected to prevent a decrease in recognition accuracy for a character string. However, the method for character recognition disclosed in Patent Document 3 fundamentally assumes that character intervals and character width are within specified ranges. Therefore, the method cannot cope with an apparatus for character recognition applying manual transfer operation in which the medium transfer speed changes on a large scale at random so that character segmentation becomes uncertain and ensuring a high recognition accuracy is difficult.